A Deep Learning Approach for Time Series Data Correction:Integrating Autoencoder-Based GANs andCorrelation Analysis

This paper presents a novel correction framework for multivariate time series data ‎that enhances data quality by integrating Autoencoder-based Generative Adversar‎ial Networks (GANs) with correlation-aware analysis. With the widespread adop‎tion of deep learning models for sensor-driven applications such as quality control ‎and demand forecasting, data anomalies caused by sensor faults and network errors ‎have emerged as a critical challenge, often degrading model performance. In mul‎tivariate time series, anomaly correction is particularly difficult because erroneous ‎values must be distinguished from legitimate temporal variations while preserving ‎inter-variable dependencies and temporal consistency. To address this challenge, ‎we propose a data correction method that combines deep learning–based anomaly ‎detection with correlation-driven correction. An Autoencoder-based GAN is em‎ployed to identify anomalous patterns in multivariate time series, while a window ‎relevance matrix is introduced to guide precise correction. This matrix captures ‎complex relationships among variables by jointly incorporating dynamic time ‎warping and Pearson correlation coefficients within sliding windows. By leverag‎ing both temporal alignment and statistical dependency, the proposed approach ‎performs anomaly correction that maintains structural coherence across variables. ‎Extensive experiments conducted on diverse multivariate time series datasets ‎demonstrate that the corrected data consistently improves predictive performance ‎compared to raw data. These results indicate that the proposed method effectively ‎enhances data quality and model reliability, offering a robust solution for anomaly ‎correction in multivariate time series applications‎.